Matt Reed
Education
M.S. Computer Science, Artificial Intelligence Concentration
Stanford University
B.S. Symbolic Systems
Stanford University
Experience
Founding Engineer
Goblins Math
- Designed knowledge-tracing and prerequisite-graph systems for adaptive math tutoring, calibrating proficiency estimates across nearly 10M student-skill observations
- Deployed self-hosted real-time tutor-avatar video streaming, reducing p99 TTFB below 1s at ~$0.016/min video cost
- Built tutor-evaluation loops to regression-test pedagogy, prompt changes, and conversation outcomes
- Developed a sandboxed SVG-generation runtime for interactive math explanations, enabling real-time visual lessons
Founding AI Engineer
Everstar
- Built chat and draft-generation systems for NRC regulatory workflows grounded in ADAMS, 10 CFR, and NUREG corpora
- Generated subsequent-license-renewal draft sections by decomposing NRC-accepted filings into reusable requirements, templates, source passages, and validation checks, reducing draft time by 10x
- Improved regulatory retrieval with contextualized chunks and sparse/dense ranking, reaching 0.815 MRR@10 and 71.4% accuracy@1 on a nuclear QA benchmark
- Built a 2,000+ example evaluation suite to test retrieval accuracy, citation coverage, source faithfulness, and draft-generation regressions
AI Researcher
Exo Labs
- Designed and ran quantitative experiments for DiLoCo + SPARTA, showing sparse asynchronous parameter averaging could increase full-sync intervals from H=100 to H=10,000 with no added wall-clock time and 14.3% lower validation perplexity
- Measured worker divergence under low-bandwidth training by tracking pairwise parameter correlation, convergence, communication overlap, and learning-rate sensitivity across SPARTA, WASH Parallel, and DiLoCo variants
- Built the initial EXO Gym emulator for running M virtual distributed-training workers on N physical accelerators, enabling experiments on single- and multi-GPU machines
- Co-authored "Improving the Efficiency of Distributed Training using Sparse Parameter Averaging" (MCDC @ ICLR 2025) and "EXO Gym" (CODEML @ ICML 2025)
Generalist Engineer
Simplify Jobs
- Implemented AI-enhanced job-listing ingestion that parsed employer websites into structured listings beyond deterministic scrapers
- Automated public internship and new-grad job boards and built promotional sites for Simplify products
- Shipped production TypeScript/React features across job search, onboarding, and marketing surfaces for student recruiting workflows
Projects
Language Model From Scratch
- Implemented a GPT-style language-model stack from scratch: byte-level BPE tokenizer, decoder-only Transformer variants, AdamW training loop, mixed precision, and Triton RMSNorm kernels
- Benchmarked training systems up to 2.7B-parameter model shapes, including PyTorch profiling, memory traces, NCCL/Gloo all-reduce, DDP gradient bucketing/overlap, and optimizer-state sharding
- Ran empirical scaling, data-quality, and alignment experiments, including OpenWebText/Common Crawl filtering, SFT, and DPO
Predicting Soil Carbon Content
- Modeled soil organic carbon using the Harmonized World Soil Database, comparing linear and nonlinear predictors under missing covariates with imputation, dropout regularization, and high-carbon-sample weighting
- Reduced weighted test MSE from 5.52 with LASSO to 0.69 with a nonlinear model, prioritizing accuracy on high-carbon soils relevant to conservation and carbon-preservation decisions
MidiVision
- Built a real-time instrument-control system using facial gesture recognition and user-trainable neural networks to map eyebrow, eye, mouth, and head movement to MIDI controls
- Designed a low-latency calibration flow that let users train personalized gesture mappings without writing code
Relevant Coursework
Ethics, Public Policy, and Technological Change/Machine Learning/Language Modeling From Scratch/Deep Generative Models/NLP with Deep Learning/Decision Making Under Uncertainty/AI Principles and Techniques/Interactive Computer Graphics/Web Applications/Computer Networking
Skills
Languages: TypeScript, Python, C++, SQL
AI/ML: PyTorch, Triton, LLM evaluation, Distributed training, Retrieval evaluation
Web/Infra: React, Effect, Postgres, OpenTelemetry, Docker
Other: Scientific writing, Data visualization, Experiment design, Product engineering